The document discusses utilizing temporal information in topic detection and tracking. It describes recognizing and formalizing temporal expressions to measure the temporal similarity between documents. The authors present their approach to resolving temporal expressions, recognizing terms, formalizing expressions as intervals on a timeline, and comparing temporal references between documents. They also discuss experiments on a news corpus and directions for future work, including improving composite expression processing and introducing vagueness.

1. University of Helsinki Department of Computer Scien
Utilizing Temporal Information in
Topic Detection and Tracking
Juha Makkonen and Helena Ahonen–Myka
{jamakkon,hahonen}@cs.helsinki.fi
University of Helsinki – Department of Computer Science
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.1/15 2003-08-1

2. University of Helsinki Department of Computer Scien
Outline
Introduction
Topic Detection and Tracking
Resolving temporal expressions
Recognition
Formalization
Comparison
Experiments
Future Work
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.2/15 2003-08-1

3. University of Helsinki Department of Computer Scien
Introduction
Temporal expressions are often omitted.
their extraction requires tools,
they have to be formalized in order to be of any use,
comparing formalizations is sometimes tricky.
By no means a novel idea
in AI to form chronologies of events,
in question answering to extract a fact,
in databases, diagnosing systems, dialog systems . . .
We want to measure the temporal similarity of two
documents.
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.3/15 2003-08-1

4. University of Helsinki Department of Computer Scien
Topic Detection and Tracking
TDT system monitors news broadcasts in order to
detect new, previously unreported events, and to
track the development of the detected events.
The focus is on news events: something untrivial taking
place at a specific time and place.
A topic is understood as as is an event or an activity,
along with all related events and activities.
The news stream that is monitored in intrinsically
sensitive to time.
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.4/15 2003-08-1

5. University of Helsinki Department of Computer Scien
Resolving Temporal Expressions
An expression can be
explicit: “the 19th of August 2003”,
implicit: “today”, “Tuesday afternoon”, or
vague: “since April”, “a couple of weeks ago” .
The evaluation is based on a point of reference. “The
winter of 1974 was cold. The next winter will be colder.”
“The winter of 1974 was cold. The next winter was colder.”
Resolving the meaning of the latter winter requires
the reference time or the utterance time and
the tense of the relevant verb.
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.5/15 2003-08-1

6. University of Helsinki Department of Computer Scien
Recognition
The relevant terms are split into categories.
category terms
baseterm day, week, weekday, month, monthname, quarter, season, year, decade
indexical yesterday, today, tomorrow
internal beginning, end, early, late, middle
determiner this, last, next, previous, the
temporal in, on, by, during, after, until, since, before, later
postmodifier of, to
numeral one, two, . . .
ordinal first, second, . . .
adverb ago
meta throughout
vague some, few, several
recurrence every, per
source from
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.6/15 2003-08-1

7. University of Helsinki Department of Computer Scien
Recognition
The categories are used to build automata.
postmodifier
determiner
ordinal postmodifier determiner
year
determiner
init
monthname
determiner
internal
temporal
numeral
internal ordinal
“The strike started on the 15th of May 1919. It lasted until
the end of June, although there was still turmoil in late
January next year”.
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.7/15 2003-08-1

8. University of Helsinki Department of Computer Scien
Formalization
We map the expressions onto a calendar
a time-line – points with precedence relation,
a set of granularities (year, month, week, . . . )
note: March, Thursday and weekend are also granularities.
a set of conversion functions between granularities.
The expressions are mapped as intervals [tstart , tend ] of the
bottom granularity which in our case is day.
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.8/15 2003-08-1

9. University of Helsinki Department of Computer Scien
Formalization
The baseterm of the expression defines interval.
The non-baseterms are interpreted as shift and span
functions that modify the start and end points.
shift: this, next, last, 3 weeks ago, etc.
span: until, before, after, from, etc.
the length of the interval is modified by internals
in the beginning of 1970s, late May, etc.
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.9/15 2003-08-1

10. University of Helsinki Department of Computer Scien
Comparison
We want to measure the temporal similarity of two
documents, i.e., how much the references overlap.
When comparing the intervals of two documents
compare pairwise all intervals
similarity = 2 * overlap / size of the intervals
take the average of the best matches for each interval.
The outcome measures how well the references of one
document cover those of the other.
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.10/15 2003-08-1

11. University of Helsinki Department of Computer Scien
Experiments
Data: transcribed TV and radio broadcasts and online
news.
8595 documents from the TDT2 corpus.
2383 documents were labeled to one of 35 events.
Temporal expression recognition with 1417 sentences
type freq recognition canonization
simple 326 0.98 0.93
composite 209 0.85 0.66
Verbs like to schedule , to plan or to expect gave hard time.
“The meeting was scheduled for Monday.” Which one?
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.11/15 2003-08-1

12. University of Helsinki Department of Computer Scien
Experiments
The distribution of temporal relations
same event
relation yes no
before 0.761 0.831
meets 0.001 0.000
overlaps 0.016 0.008
begins 0.010 0.006
falls within 0.168 0.122
finishes 0.010 0.008
exact 0.072 0.056
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.12/15 2003-08-1

13. University of Helsinki Department of Computer Scien
Experiments
Temporal similarity is higher when documents are
relevant.
average of same event different event ratio of yes/no
sum of pairwise 0.0034 0.0023 1.4783
max of pairwise 0.0059 0.0040 1.4750
Finding the best-match for each interval does not pay off.
A better accuracy on formalization would help.
What is the meaning of “three years ago?”
How to represent informativeness?
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.13/15 2003-08-1

14. University of Helsinki Department of Computer Scien
Future Work
Improvement of the composite expression processing
more work on the automata
Introduction of vagueness:
an expression would formalized as probability
distributions on the timeline
similarity could be Kullback-Leibler, for instance.
Survey of the behaviour of the temporal expressions
how the references distribute per medium?
the first story compared to the following ones?
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.14/15 2003-08-1

15. University of Helsinki Department of Computer Scien
The End
Thank you
Juha Makkonen and Helena Ahonen–Myka Utilizing Temporal Information in Topic Detection and Tracking – p.15/15 2003-08-1